Comment by sandpaper26

It's not a likely solution given how loss functions work, but in theory a single model could learn to perform exactly the function you describe. When you say "just do X" where X is any function (in this case, a piecewise function), a large enough model could do it.